--- title: How I built Wordle, but it's Tetris description: "I tested a cheap AI model pipeline by building Swipetris, a daily Tetris challenge designed to be more fun than another feed." pubDate: "2026-08-20" --- When I read about the release of [DeepSeek-V4-Flash](https://www.deepseek.com/en) and its comparisons with [Opus](https://www.anthropic.com/claude/opus), I was intrigued... and a bit incredulous. A model that's [10-20 times cheaper](https://artificialanalysis.ai/models/deepseek-v4-flash) with very similar results? I had to check it out. Also, trust-me-bro benchmarks have to be taken with a grain of salt anyway. There was also the [Kimi K3](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) release around the same time, so I wondered: what if I used both to build something and compare to the big ones, like Sol or Fable? Now I just needed something to build with... So I looked around me, and saw the same thing happening everywhere, **doomscrolling**. Instagram, TikTok, etc... when people have a minute to spare, they usually are swiping around, getting shots of distraction while time passes. Am I innocent? No, I'm not. Reddit, blog posts or nerdy rabbit holes are my thing, not so much as swiping up every 3 seconds, I prefer a bit longer interactions, but I digress... ### So, swiping is inevitable... If it is, why not use the same gesture for something fun and free instead of another feed? Being an 80's kid, what came to mind was [Tetris](https://tetris.com/). But with a twist. **A daily challenge, with the same block sequence for everyone**. You can play as many times as you want and fix your mistakes, go further and only your highest score will make it onto the leaderboard. ## I can just download Tetris True, and I did actually as a benchmark. Then came the subscription prompt, coins, ads after each level, location tracking... for Tetris??? I wouldn't want to improve what was there. I wanted to build a version I'd like to exist. ## Another world's favourite: Wordle [Wordle](https://www.nytimes.com/games/wordle/index.html) was created as a gift, and the catch is that you have one puzzle a day, same for everyone. Scores are then comparable. That was the other half of the idea: **one shared challenge a day**. Everyone has the same puzzle to solve, so scores mean something. ## Plan done, now the models I had the idea. Now I gotta be careful not to [keep spinning my wheels](/blog/planned-28-apps-shipped-2) and get down to action. It was time to see what these models could do, especially DeepSeek-V4-Flash. ### One thing was immediate: it is **fast** Actually, two things were impressive: - How fast it churns tokens and - how good the results are. But planning with it was a bit... meh. It sounded like it was making an effort, but it was not made for this. It reminded me of Sonnet 4. Then I thought I'd reserve it for execution, and turned to Kimi K3 for architecture and planning. Where I spend most of the time, anyway. I also remembered that K3 had good feedback on three.js, so that part is on it. Surprise, K3 is also quite good at design! And it's multi-modal, meaning I can just paste screenshots and use it basically like I'd use Sol or Fable. After Kimi was done with planning and the three.js part, DeepSeek-V4-Flash took care of all the other tasks. Many small tasks, and at the end of it, I looked at the cost, and there was barely any. Less than a dollar for multiple rounds and tasks. DeepSeek-V4-Flash then was used for controls, PWA setup, score logic, the bot (the noon leaderboard cheeky one and the actual player bot that doesn't save scores) and fixes. Once you have a solid base, it can do most tasks. At the end, I put on Fable for final touches in the design, because of its taste. However further developments were all done with DeepSeek-V4-Flash. My conclusion is that **a pipeline beats a single model**. Both in cost and effectiveness. Plan with K3, execution with DeepSeek-V4-Flash, polish with Fable or by hand. ## Launch I just announced on Reddit, where I got a few "it's just another Tetris", which was partly the point. ![Swipetris landing page: play now, no install, swipe to move](/swipetris-screenshot.webp) But then some people started to come back. And some of them (FRED) are absurdly good! Seeing scores 3 times higher than my best on a thing I built is a nice humbling experience. ## Technical details This is a PWA (Progressive Web App), that you can add to your home screen and behaves like a native app. Super light, works offline, no tracking of any kind, fully open source. ## Main takeaways [Swipetris](https://swipetris.com) started as an excuse to test those models on something real and see how capable they are. And have fun doing it. And fun it is. To build, to play, to add features. The best outcome is seeing people I don’t know come back and beat my scores. I really believe in creating things and releasing for fun. I’ve built my career as a web designer/developer with open-source software, from the tools I use every day to the servers and infra behind what I ship. So creating things and giving them away for free, in my humble opinion is a way to give back.